Marketing Applications of ChatGPT: User Interviews

It has been exactly one month since OpenAI released the high-performance AI chat,ChatGPT, on November 30th. Since its release, it has garnered attention for its high performance, and various use cases have been published.

I have tried various things as ideas for marketing applications.

The previous article on "Inquiry Support" can be found here.

This time, I will take on the challenge of "User Interviews."

Regarding similar services

Actually, Dentsu offers a similar service.

This service uses "Kiku-Hana," a chatbot engine independently developed by Dentsu. It is specialized for business use, and its algorithm is built around a rule-based system to ensure it does not lie or provide false information.

However, this time I will replace this with ChatGPT.

1. Digging into purchasing motives

When considering marketing for existing products, we often look at the purchasing motives of current users. This is because the features that resonate with them may differ from the strengths assumed by the provider, or the product may be used in ways that were not anticipated.

As a research method, there is analysis of social media and reviews, but to understand things more deeply, we often conduct specialized user research.

With conventional methods, it would cost several hundred thousand yen for research design, moderator fees, and analysis fees (and you could only survey a few people), but this time I will use ChatGPT to substitute for the interviewer and the report creation process.

As for the subject, I purchased a Pixel 7 smartphone last year. This time, I will have it dig into the purchasing motives for that Pixel 7. Since smartphones are products that involve comparative consideration, I think they fall into a category that is quite easy to interview about.

Giving instructions

I gave instructions like this.

This is all that is needed.

It is important to include the prompt, "Keep questions concise and ask only one question at a time," otherwise it will become a chaotic interviewer that bombards you with about 10 questions at once.

And so, it has begun.

I'm incredibly excited.

Answering

I'll keep answering from here on out. Since it's about me, it's easy to answer.

The interview is proceeding normally. For some reason, it asks two questions at a time, but I'll just keep going without worrying about it.

It's asking all the standard questions that usually come up in interview surveys, like "comparison with other products" and "points of praise and dissatisfaction." I can feel its proficiency.

This guy is good...

It's extracting items that are usually asked in standard marketing research, such as information sources and consideration routes, one after another.

I answered the questions quickly, but then an unexpected development occurred.

The last question was a bit playful, overlapping with previous ones, but more than that, this little one remembered it was a roleplay and ended the interview at the appropriate time. How amazing!

Finally, I made this request.

And then,

It turned out to be a much more proper interview than I imagined, and I couldn't help but exclaim, 'Wow!'

Verify further

Assuming I ask many users the same questions, I tried it again from the beginning using the exact same prompt.

The interviewer this time is good at providing feedback. It's nice that the interviewer's characteristics are different each time, which feels human.

Amazing...

(2) Digging into insights

It's relatively easy to ask about specific products, but this time I'll try something a bit more abstract: 'digging into what users think about a certain theme (insights).'

This time, I don't own a car, so I'll have it dig into the reasons why I don't own one.

Give instructions

Answer

The next interviewer has a bit of a quirk with their backchanneling. Let's take a look.

The questions are becoming a bit harder to answer. It might have been a mistake to use abstract words like "consciousness" in the initial prompt. Also, I feel like it's talking a bit too much in general.

Trying to compensate for a lack of knowledge or confidence with verbosity is human-like, but it's not great for an interview.

so it seems it's not very good at digging into abstract insights after all.

Since it was like this the whole time, I stopped the interview halfway through.

Improve

However, I really want it to conduct a proper interview, so I will try a different technique.

As I wrote in a previous entry, ChatGPT is basically designed based on English, and it only learned Japanese as "incidental training data." Therefore, the overall accuracy is higher in English.

I feel like it's conducting the interview much more smoothly than when I do it in Japanese. It gives the impression of thoroughly filling in peripheral information, such as past history and current means of transportation.

As an aside, the response is also much faster in English. It's easy to see why there's a ChatGPT trick of 'having it translate to English, answer, and then translate back to Japanese.'

Let's keep moving forward with the answers.

It asked about a wide range of things quite smoothly. Moreover, the questions were very simple and easy to answer. It's quite different from when I did it in Japanese...!

And I had it output the summary in Japanese.

My last statement, 'summery' isn't a verb so it's a grammatical error, but it takes that in stride with ease.

I feel like there are some issues with the final Japanese translation, but overall it's well done.

So, as is often said, I've found that the accuracy is higher in English.

Summary

I was able to have ChatGPT conduct an interview survey. Honestly, I was surprised at how much better it was at interviewing than I expected. My feeling is that it is already sufficiently usable at this point if you just want to flip through the results and perform a qualitative analysis.

The final summary items will likely differ depending on the survey respondent, but if there were a way to effectively consolidate and organize them, it would be possible to perform quantitative analysis as well. Alternatively, if I refine the initial prompt, it might be possible to extract the same items.

That's all for today.